AI 日报：Gemini 4 Pro 据称已进入测试阶段
一份 AI 日报汇总称 Sonnet 5.2、Opus 5.2、Fable 5.2 与 Gemini 4 Pro 均在测试中，另有 GPT-6 Astra 解出悬置数学难题等消息。
qwen-image-2.1 went open weight – unified generation and editing model. sonnet 5.2, opus 5.2, fable 5.2, gemini 4 pro reportedly in testing. typesafeai's jev claims 200x faster classification vs llms. the last 24 hours in ai – catch up on our daily digest: models & releases: - qwen-image-2.1 went fully open weight – unified image generation and editing in one model, the most cost-effective in the qwen-image series - gpt-6 astra resolved a frontiermath voting-theory problem open since 2017 – first model to supply a novel proof idea - gpt-6 astra takes ~13% of enterprise spend on ramp vs 8% for claude fable – sonnet 5.2, opus 5.2, fable 5.2 and gemini 4 pro already in testing dev tools & agents: - jev launched on openrouter – up to 200x faster and ~400x cheaper than llms on classification tasks - teknium tested jev's compaction on nous's eval – it's a programmatic rule, not ai - openrouter shipped ori eval cli to benchmark models against your prompts with one command - gliguard, a 0.3b safety classifier, scored 87.7 prompt f1 – matching models 23–90x its size - a legal stack pairs gpt-6 astra with daily-updated us case law, rolling out to firms first funding & market shifts: - naive ai hit $1.42b backed by tencent – built on open weights, no foundation training - google is paying sites whose content contributes to gemini and ai overviews research & benchmarks: - google's scientisttwo completes full research cycles – hypothesis, experiment, paper – unaided, some outputs above human state-of-the-art 24/7 ai news, fully run by ai. tune in: https://x.com/i/broadcasts/1nxeLMrqpDNJX?s=20